https://orcid.org/0000-0001-9384-4461
Figures
Abstract
Soil erosion is one of the most widespread environmental issues globally, posing serious threats to ecosystems and land resources. This study employs precipitation, soil, digital elevation model, and land-use data from 2000 to 2020 to quantitatively analyze the spatiotemporal patterns of land-use change and soil erosion in the Kangding River Basin through GIS-based spatial analysis and the RUSLE (Revised Universal Soil Loss Equation) model, and to evaluate soil stability across the watershed. Furthermore, using Geographical Detector (including single-factor detection and dual-factor interaction detection) and the SHAP (SHapley Additive exPlanations) algorithm to analyze the optimal machine learning model enables the assessment of the contribution of each driving factor to soil erosion. The results revealed that: (1) From 2000 to 2020, the areas of woodland and water body exhibited a decreasing trend, while cropland and construction land expanded steadily.(2) The soil erosion modulus in the Kangding River Basin first increased and then declined during the study period, rising from 16.32 t·hm-2·a-1 in 2000 to a peak of 21.85 t·hm-2·a-1 in 2005, and subsequently decreasing to 11.16 t·hm-2·a-1 by 2020.(3) The predominant erosion intensity was slight and very slight, with an increase in erosion severity between 2000 and 2005, followed by a gradual decrease thereafter.(4) The results of the geographical detector and SHAP analysis indicate that slope, land use, and vegetation coverage were the three most influential driving factors affecting soil erosion in the basin. These findings provide a scientific basis for comprehensive watershed management and land use planning in the Kangding River Basin, offering important theoretical support for soil and water conservation in the region.
Citation: Guan Y, Duan X, Deng Q, Chen B, Su B, Zeng K (2026) Spatiotemporal characteristics and driving factors of soil erosion in the Kangding River Basin (Southwest China) based on the RUSLE model. PLoS One 21(3): e0344489. https://doi.org/10.1371/journal.pone.0344489
Editor: Claudionor Ribeiro da Silva, Universidade Federal de Uberlandia, BRAZIL
Received: August 25, 2025; Accepted: February 21, 2026; Published: March 16, 2026
Copyright: © 2026 Guan et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The data used in this study include: Sentinel-2 imagery for the study area, downloaded from the Copernicus Open Access Hub (https://dataspace.copernicus.eu, accessed on 25 May 2025) and used for spatial analysis and visualization; precipitation data from the Resource and Environment Data Center, Chinese Academy of Sciences (https://www.resdc.cn/data.aspx?DATAID=379, accessed on 25 May 2025); land use/land cover (LULC) data from the CNLUCC (China’s Multi-Period Land Use Land Cover Remote Sensing Monitoring Dataset), with a spatial resolution of 30 m (https://www.resdc.cn/DOI/DOI.aspx?DOIID=54, accessed on 25 May 2025); digital elevation model (DEM) data from the Geospatial Data Cloud (https://www.gscloud.cn/sources/accessdata/310?pid=302, accessed on 25 May 2025); and soil erodibility factor data derived from the 1:1,000,000 World Soil Dataset (https://www.fao.org/soils-portal/data-hub/soil-maps-and-databases/harmonized-world-soil-database-v12/en/, accessed on 25 May 2025). The normalized difference vegetation index (NDVI) data were obtained from the Resource and Environmental Science Data Platform, with a spatial resolution of 30 m (https://www.resdc.cn/DOI/DOI.aspx?DOIID=68, accessed on 25 May 2025). Slope and slope direction data were derived from the DEM. All data used in this study were processed to have a uniform spatial resolution of 30 m. Access to RESDC datasets requires free user registration on the platform. All datasets are hosted on public data platforms, and each dataset can be accessed through the official webpage we have provided (e.g., RESDC DataID pages, CNLUCC dataset DOI pages, Geospatial Data Cloud download page, FAO soil dataset site, and the Copernicus Data Space portal). These datasets are openly accessible to all users. For RESDC datasets, users are required to create a free account before downloading. No special permission is required.
Funding: This study was financially supported by the National Natural Science Foundation of China (Grant No. 42201006) and the Natural Science Foundation of Sichuan Province (Grant No. 2022NSFSC1177).
Competing interests: The authors have declared that no competing interests exist.
1. Introduction
Soil erosion is a critical indicator for evaluating ecosystem resilience, directly influencing soil fertility, and water conservation, ecological stability [1,2]. A healthy soil environment possesses strong capacities for precipitation interception, water conservation, and ecological support, forming the foundation for maintaining biodiversity and the stable functioning of ecosystems [3]. Accelerated anthropogenic activities (e.g., irrational land development and vegetation destruction) and natural factors (e.g., precipitation variability, topographic relief), have resulted in land degradation, increased river sedimentation, and diminishing habitat sustainability [4,5]. Consequently, a clear understanding of the interactions between land use changes and natural factors is essential for enhancing soil and water resource management as well as developing effective conservation strategies [6,7].
Existing studies have explored the driving factors and response mechanisms of soil erosion from multiple perspectives. Pham et al. systematically assessed the land use transformation characteristics in rapidly developing watersheds using remote sensing and GIS technologies, and found that the significant expansion of construction land directly contributed to the increase in soil erosion modulus [8]. Wen et al., through a global-scale meta-analysis, identified land use fragmentation as a key factor driving the spatial heterogeneity of soil erosion [9]. In sloped environments, agricultural development has been confirmed as an important driver of severe soil erosion. Based on historical data, Chaudhary and his team revealed that large-scale cultivation on slopes induces serious soil loss and ecological degradation. Irrational farming practices not only destroy surface vegetation but also alter soil structure and reduce its resistance to erosion [10].
In recent years, emerging studies have provided deeper insights into soil erosion, with research shifting from static assessments to dynamic simulations and multi-scenario forecasting. Wu et al. assessed soil erosion changes in the Yellow River Basin from 2010 to 2020 based on time-series remote sensing data, and found that ecological engineering projects significantly curbed erosion expansion after 2010, although localized human activities continued to pose risks [11]. She analyzed the spatiotemporal evolution of erosion in the Weihe River Basin from 2000 to 2020 and revealed its coupling relationship with land development intensity [12]. Jiao et al. conducted multi-scenario simulations to predict erosion trends under various land use strategies in a typical karst basin, providing a basis for ecological restoration [13]. Similar patterns have also been observed in international studies. Arunrat et al. reported that erosion modulus in agricultural areas of northern Thailand was significantly higher than in areas with natural vegetation [14]. Mokarram et al., in southwestern Iran, combined UAV-based remote sensing and field soil moisture monitoring to uncover the feedback mechanisms between vegetation physiological responses and soil erosion [15]. Seitkazy et al. employed a convolutional neural network integrated with the RUSLE model to analyze the impacts of topography (LS factor) and farmland erosion in western Kazakhstan, and predicted erosion risks under different climate and land use scenarios [16]. Waseem et al. evaluated the influence of urban and agricultural development activities on soil erosion risk in the Mangla Reservoir, identifying impervious surfaces and slope as key factors contributing to erosion threats in built-up areas [17].
As a major tributary of the upper Yangtze River, the Kangding River Basin features complex terrain, variable climate, and a fragile ecosystem, with soil erosion emerging as a particularly prominent issue. Compared with other Yangtze tributaries such as the Minjiang and Jialing Rivers—which have been extensively studied in terms of land-use change, hydrological regulation, and soil conservation measures [18,19]—the Kangding River Basin has received relatively limited scientific attention. Despite sharing similar ecological vulnerability and experiencing rapid socio-economic transformation, this persistent research gap underscores the regional novelty and significance of focusing on this under-studied mountainous watershed. In recent years, the soil erosion processes in the basin have undergone significant changes due to the combined impacts of economic development, tourism expansion, and climate change in western Sichuan. However, existing studies still exhibit the following limitations: (1) a lack of in-depth analysis of the spatial heterogeneity of soil erosion under the integrated terrain–climate–land use system [20]; and (2) limited attention to multi-scenario simulations and policy intervention assessments in previous research, which has constrained the scientific support for watershed ecological management [21]. While our study does not explicitly conduct future scenario simulations, identifying this gap highlights a key direction for subsequent research.
This study focuses on the Kangding River Basin and adopts an integrated approach combining remote sensing monitoring and spatial analysis to conduct systematic research in three main aspects: (1) analyzing the spatiotemporal characteristics of land use changes over the past two decades; (2) assessing soil erosion intensity and its spatial distribution patterns across different years using the RUSLE (Revised Universal Soil Loss Equation) model; and (3) identifying the primary driving factors of soil erosion and their interaction effects through the application of the geographical detector, machine learning models and SHAP (SHapley Additive exPlanations) analysis. The aim is to uncover the spatial heterogeneity and driving mechanisms of soil erosion in typical mountainous watersheds, thereby providing scientific support for regional ecological protection and soil and water conservation planning.
2. Overview of the study area
The Kangding River Basin, with a total area of 210.2 km², is located in Kangding City, Sichuan Province, China, spanning from 101°57′ to 102°10′E and 30°01′ to 30°10′N. The basin lies at elevations ranging from 1,333–5,631 meters and is characterized by an average annual precipitation of 845.4 mm, with rainfall concentrated in specific periods. The mean annual temperature ranges from 3.5°C to 7.1°C, and the region is classified as having a subtropical monsoon climate (Fig 1). The Kangding River, a major tributary of the Dadu River, originates in Dapaoshan Mountain (with a peak elevation of approximately 4,000 m) in the northern part of Kangding City, Ganzi Tibetan Autonomous Prefecture, Sichuan Province. The river is primarily sourced from the Yala River and flows into the Dadu River at Wasi Valley in the eastern urban area of Kangding City. It mainly flows through deeply incised valleys, exhibiting a dendritic drainage pattern with an average channel slope of approximately 10%. The annual runoff volume is about 1.5 × 10⁸ m³, and the annual sediment transport rate is around 50 m³/s. The theoretical hydropower potential of the basin is estimated at 2.0 × 10⁴ kW. Hydraulic erosion is the dominant type of soil erosion in the basin, while gravitational and aeolian erosion occur in limited areas.
3. Materials and methods
3.1. Data sources and data processing
The data used in this study (Table 1) include: Sentinel-2 imagery for the study area, downloaded from the Copernicus Open Access Hub (https://dataspace.copernicus.eu, accessed on 25 May 2025) and used for spatial analysis and visualization; precipitation data from the Resource and Environment Data Center, Chinese Academy of Sciences (https://www.resdc.cn/data.aspx?DATAID=379, accessed on 25 May 2025); land use/land cover (LULC) data from the CNLUCC (China’s Multi-Period Land Use Land Cover Remote Sensing Monitoring Dataset), with a spatial resolution of 30 m (https://www.resdc.cn/DOI/DOI.aspx?DOIID=54, accessed on 25 May 2025); digital elevation model (DEM) data from the Geospatial Data Cloud (https://www.gscloud.cn/sources/accessdata/310?pid=302, accessed on 25 May 2025); and soil erodibility factor data derived from the 1:1,000,000 World Soil Dataset (https://www.fao.org/soils-portal/data-hub/soil-maps-and-databases/harmonized-world-soil-database-v12/en/, accessed on 25 May 2025). The normalized difference vegetation index (NDVI) data were obtained from the Resource and Environmental Science Data Platform, with a spatial resolution of 30 m (https://www.resdc.cn/DOI/DOI.aspx?DOIID=68, accessed on 25 May 2025). Slope and slope direction data were derived from the DEM. All data used in this study were processed to have a uniform spatial resolution of 30 m. Access to RESDC datasets requires free user registration on the platform.
3.2. Methods
The research methodology mainly includes the following (Fig 2): First, the RUSLE model was applied to simulate soil erosion in the study area. Second, the Geographical Detector method was used to analyze the driving factors of soil erosion. Third, SHAP was employed to interpret the optimal machine learning model.
3.2.1. RUSLE model.
The Revised Universal Soil Loss Equation (RUSLE) was applied to estimate soil erosion in the study area [22]. While precise estimation is affected by uncertainties in rainfall erosivity, soil properties, and topographic parameters, previous case studies under mountainous terrain suggest that RUSLE-derived soil erosion estimates generally carry an uncertainty of approximately ±10–20% [23,24]. Its mathematical expression is:
in the equation, A represents the soil erosion modulus (t·hm-2·a-1); R is the rainfall erosivity factor (MJ·mm·hm ⁻ ²·h ⁻ ¹·a ⁻ ¹); K denotes the soil erodibility factor (t·MJ ⁻ ¹·mm ⁻ ¹·h); L and S are the slope length and slope steepness factors, respectively; C is the cover and management factor; and P is the support practice factor.
Since R is difficult to measure directly, it is typically estimated using parameters such as rainfall amount and intensity. The R factor quantifies the amount and rate of rainfall impact on runoff generation [25,26]. In this study, the monthly R values were calculated using the Wischmeier empirical formula based on the multi-year average monthly precipitation [27]:
where and p represent the average monthly precipitation and the average annual precipitation (mm), respectively.
The monthly precipitation values used in the formula were derived from daily rainfall data (≥10 mm) [28], which accounts for daily rainfall intensity and event distribution. This approach ensures that variations in rainfall strength and duration are captured while maintaining comparability with previous studies [29].
The soil erodibility factor (K) describes the susceptibility of soil to erosion under a given rainfall erosivity. The K value can be determined through soil physical and chemical properties or by hydraulic erosion experiments [30–32]. The magnitude of K is closely related to soil organic matter content and soil texture. In this study, a calculation formula proposed by Zhang Keli, which is better suited to the characteristics of Chinese soils, was adopted [33]. This formula takes into account multiple soil properties, including sand, silt, and clay contents, as well as soil organic carbon content. The calculation formula is as follows:
where is the sand content (%),
is the silt content (%), n is the clay content (%), and C is the soil organic carbon content (%);
. The unit of K is in the U.S. customary system and should be converted to the International System of Units (SI) by multiplying with a conversion factor of 0.1317. This conversion was already applied in all calculations, and an explicit note has been added here for clarity.
The slope length and steepness factors (LS) were derived using digital terrain analysis techniques. Currently, the LS factor is calculated using digital terrain analysis techniques; however, its accuracy is limited by the quality and type of DEM data. In this study, multi-source DEM data were integrated for improved calculation. First, a zonal weighted fusion of the multi-source DEM data was performed. Based on the fused regional DEM elevation map, slope length and steepness information were extracted, which significantly enhanced the accuracy of the LS factor [26]. The specific formula is as follows:
where L represents the soil erosion amount normalized to a slope length of 22.13 m, and λ denotes the slope length.
The slope steepness factor (S) is calculated using a segmented approach [22], and its expression is given as follows:
where S represents the dimensionless slope steepness factor, and θ denotes the slope angle (in degrees, °) measured from the horizontal plane.
The cover and management factor (C) represents the effect of vegetation cover and land use on soil erosion and ranges from 0 to 1 [34]. According to Molla, Ganasri, and others [35,36], the C factor quantifies the cumulative effect of trees, crop sequences, and other land cover types on land degradation. It is closely related to vegetation cover and land use, with values ranging from 0 to 1. In this study, the formula proposed by Cai was adopted [37], and the expression for calculating C is as follows:
where the expression for fc is given as follows:
in this expression, denotes vegetation cover (%);
is the normalized difference vegetation index.
and
represent the minimum and maximum NDVI values in the study area, respectively, with the 5th and 95th percentiles selected as the minimum and maximum thresholds.
The support practice factor (P) represents soil and water conservation measures and is defined as the ratio of soil loss after implementing conservation practices to the soil loss from contour farming. Its value typically ranges from 0 to 1, where 1 indicates areas without any conservation measures, and 0 corresponds to areas where erosion does not occur [38]. To build upon previous studies (Table 2), values were assigned in combination with practical conditions [39,40].
3.2.2. Classification of soil erosion intensity.
Soil erosion intensity is a critical indicator for assessing the degree and rate of soil erosion, providing a scientific basis for evaluation, management, and monitoring of soil erosion processes [42,43]. According to the “Soil Erosion Classification and Grading Standard” (SL190–2007) issued by the Ministry of Water Resources of China in 2008, the soil erosion modulus simulated by the RUSLE model within the study area is classified as shown in Table 3 [44].
3.2.3. Assessment of soil stability.
Soil stability assessment is of great significance for the rational planning of land use, ecological protection, soil and water conservation, and the safety of infrastructure. It serves as a fundamental task for achieving sustainable development [45]. In this study, the soil stability of the Kangding River Basin was analyzed and evaluated based on DEM data and land use maps (Table 4), and the corresponding soil stability factors were assigned and calculated accordingly [46,47].
The calculation formula after assignment is as follows:
in the equation, reland represents the reclassified land use value; reaspect denotes the reclassified aspect value; and reslope refers to the reclassified slope value.
Based on the soil stability values, the land stability in the Kangding River Basin was classified into three levels: low stability (<4), moderate stability (4–7), and high stability (>7) [48].
3.2.4. Geographical detector.
The Geographical Detector model proposed by Wang et al. [49] was applied to analyze the driving factors of soil erosion [49]. The degree of spatial differentiation is measured by the q-statistic. Since it typically uses categorical variables as explanatory factors, it is particularly suitable for detecting soil erosion patterns [50]. In this study, the factor detector and interaction detector modules were employed for quantitative analysis of the study area. All raster datasets were resampled to a spatial resolution of 30 × 30 m. Continuous variables were discretized into strata as follows: elevation, rainfall, slope, and land use were divided into six classes, while vegetation cover was divided into five classes.
The factor detector was used to quantify the explanatory power of each driving factor on soil erosion. This explanatory power is measured by the q-statistic, calculated as follows:
where h = 1, 2,..., L represents the strata of the independent variable X; Nh and N denote the number of units in stratum h and in the entire study area, respectively; and
are the variances of the dependent variable Y within stratum h and the entire region, respectively. SSW is the sum of within-stratum variances, and SST is the total variance of the region. The q value ranges from 0 to 1, with higher values indicating stronger explanatory power. A q value closer to 1 indicates stronger explanatory power. A simple transformation of q follows a non-central distribution, and its statistical significance can be tested using the Geodetector software. In this study, to ensure robust significance evaluation, a permutation test with 999 iterations was additionally conducted on all q-values.
The interaction detector can also be used to identify the interaction effects between different factors. Specifically, it evaluates whether the combined influence of two factors (X₁ and X₂) on the dependent variable Y is enhanced or weakened by calculating q (X₁ ∩ X₂). The value of q(X₁ ∩ X₂) is then compared with q(X₁) and q(X₂) to determine the nature of the interaction effect [49]. The criteria for interpreting the relationship between the factors and the dependent variable Y are presented in Table 5.
3.2.5. SHAP and machine learning model.
SHAP, proposed by Lundberg and Lee (2017), is a unified framework for interpreting machine learning model predictions [51]. It quantifies the contribution of each feature to model predictions by computing marginal effects across feature combinations [52,53]. In this study, three machine learning models were used: CatBoost, XGBoost, and LightGBM.
For the machine learning models, the dataset was randomly split into a training set (70%) and a test set (30%). Hyperparameters were manually tuned through empirical evaluation to achieve stable and reliable model performance, following practices adopted in previous studies [54,55].
- 1). CatBoost model
The CatBoost model is a machine learning algorithm based on the gradient boosting framework. It is specifically optimized for handling categorical features by automatically encoding categorical variables and employing symmetric decision trees. CatBoost is particularly well-suited for datasets with a large number of categorical variables [56]. In this study, soil erosion estimates derived from the RUSLE model were used as the target variable, and the driving factors identified by the Geographical Detector were used as model inputs.
- 2). XGBoost model
XGBoost is a distributed gradient boosting framework based on decision trees and is widely used in various machine learning tasks. Its core idea is to combine simple models (weak learners) into a strong model by iteratively correcting prediction errors to continuously improve predictive performance [57]. The objective function of XGBoost consists of two components: the loss function and a regularization term [58].
- 3). LightGBM model
Compared with conventional gradient boosting models, LightGBM improves training efficiency and reduces memory consumption by adopting a histogram-based algorithm. It employs a leaf-wise growth strategy with depth constraints, enabling efficient learning of complex nonlinear relationships.
In this study, LightGBM was implemented using its standard regression setting with Mean Squared Error (MSE) as the loss function. Model performance was evaluated using statistical metrics, and SHAP was applied to interpret the contribution of driving factors to soil erosion [40].
4. Results
4.1. Land use change characteristics in the kangding river basin
In the Kangding River Basin, woodland and grassland occupy the largest areas and are primarily distributed in the mountainous regions on both sides of the river. Cropland is mainly located along the middle and lower reaches of the river. Construction land is concentrated in flat areas at both ends of the basin. From 2000 to 2020, the area of construction land in the Kangding River Basin has continuously increased, accompanied by a rise in the number of unused land patches (Fig 3).
To conduct an in-depth analysis of land use changes in the Kangding River Basin from 2000 to 2020, this study employed a land use transfer matrix to quantify conversions among different land use types. The cropland area in the basin increased from 5.73 × 10² hm² to 6.96 × 10² hm², representing a total increase of 21.41%, with its expansion primarily originating from grassland and unused land. Woodland, the dominant land use type in the basin, declined from 1.28 × 10⁴ hm² to 1.12 × 10⁴ hm², a total decrease of 11.72%. Over the four sub-periods, approximately 0.5%–1.5% of woodland was converted to grassland, and a small fraction (< 1%) was converted to unused land. Grassland area increased from 5.54 × 10³ hm² to 5.70 × 10³ hm², showing a total growth of 2.86%, mainly due to woodland degradation and the abandonment of cropland. Approximately 0.5%–2% of grassland was converted to unused land, with only a negligible portion converted to construction land. Water bodies decreased from 2.49 × 10² hm² to 2.19 × 10² hm², a total reduction of 12.04%, with both inflow and outflow remaining minimal (< 1%). Construction land expanded from 5.74 × 10² hm² to 1.18 × 10³ hm², representing a total increase of 105.44%, and was primarily concentrated in the flat areas at both ends of the basin. Unused land increased from 1.26 × 10³ hm² to 1.89 × 10³ hm², reflecting a total growth of 49.18% (Fig 4).
4.2. Spatiotemporal variation of soil erosion
From 2000 to 2020, the soil erosion modulus in the Kangding River Basin exhibited a trend of initial increase followed by a subsequent decrease (Fig 5). Overall, the spatial distribution of soil erosion modulus displayed a distinct pattern of “high values along the margins and low values in the central area”. Areas with moderate to high erosion were primarily concentrated on the steep slopes along the basin margins, particularly in the northern and southeastern regions, whereas low-erosion areas were widely distributed across the central, relatively flat zones. This spatial pattern remained generally consistent across years, although erosion intensity exhibited noticeable temporal variation.
Based on the 20-year soil erosion modulus data in the Kangding River Basin (Fig 6), the overall soil erosion modulus was relatively low in 2000, with scattered moderate to high value areas appearing only in the northern margin and southwestern regions. By 2005, the erosion modulus showed an increasing trend, with significant expansion of moderate to high value zones, particularly forming continuous belt-like distributions in the northern and southeastern areas. The erosion modulus peaked at 21.85 t·hm-2·a-1, up from 16.32 t·hm-2·a-1, marking the highest value within the 20-year period, indicating a notable intensification of erosion that year. Although high-value areas persisted in 2010, especially in the northern region, their spatial extent and continuity diminished compared to 2005, with the erosion modulus decreasing to 18.05 t· hm-2·a-1. In 2015, erosion intensity further declined, with a marked contraction of moderate to high value areas, leaving only a few high-value points visible along the southern margin. This trend continued in 2020, with the spatial pattern stabilizing and erosion modulus values of 11.33 t· hm-2·a-1 and 11.16 t· hm-2·a-1, respectively. It is important to note that the RUSLE-derived soil erosion estimates carry an uncertainty of approximately ±10–20% due to parameter error propagation.
4.3. Spatial variation of soil erosion intensity
To systematically analyze the spatial distribution and temporal variation of soil erosion intensities in the study area, the soil erosion modulus of the Kangding River Basin was reclassified in ArcMap according to the “Soil Erosion Classification and Grading Standard” (SL190–2007) since 2000, resulting in different erosion intensity categories (Fig 7). The area and percentage of each erosion intensity class were calculated for the study region (Table 6). The majority of the basin is dominated by slight, mild, and moderate erosion, followed by strong and very strong erosion, while severe erosion occupies the smallest area. The area under slight erosion decreased from 81.26% in 2000 to 76.12% in 2005, but showed a substantial increase after 2005, reaching 89.95%, indicating a transient decrease followed by a strong recovery in slight erosion coverage. Mild erosion area exhibited an increasing trend from 2000 to 2005, then began to decline thereafter. Similar to slight erosion, moderate and strong erosion areas displayed an “increase–decrease–increase” pattern, with their coverage rising from 2000 to 2005, then decreasing from 5.29% and 4.21% in 2010 to 3.03% and 1.42% in 2020, respectively. The proportion of very strong and severe erosion showed a general declining trend over the past two decades. Overall, soil erosion intensity in the Kangding River Basin has shifted from higher to lower intensity classes, although transitions between slight and mild erosion remain relatively active.
4.4. Spatial differentiation of soil stability
Fig 8 illustrates the spatial distribution and temporal variation of soil stability classes in the study area over the past 20 years. Marked differences in soil stability are evident on either side of the river, which acts as a natural boundary.
In 2000 (Fig 8a), areas of high stability were widely distributed across the central and eastern parts of the basin, exhibiting strong spatial continuity. Low stability zones were mainly scattered in the northwestern and southern margins of the river, covering relatively small areas. By 2005 (Fig 8b), the extent of low stability regions expanded significantly, especially along the western and southern edges of the river basin. In 2010 (Fig 8c), compared to 2005, high stability areas further extended east of the river, while low stability zones slightly contracted but remained concentrated in the northern part. During 2015, the moderate to high stability zones continued to expand, with a reduction and spatial restriction in low stability areas. By 2020 (Fig 8e), the overall soil stability pattern remained relatively stable; high stability areas were broadly distributed throughout the basin, whereas low stability zones were mainly confined to localized regions in the northwest.
Based on the comparison of five temporal images, soil stability in the Kangding River Basin was generally dominated by moderate to high stability classes. Low-stability areas were primarily distributed along the western side of the river and basin margins, where their extent alternated between expansion and contraction from 2000 to 2020. The temporal evolution of stability classes exhibited distinct stage-wise characteristics, with the river acting as a clear spatial boundary in the distribution pattern.
4.5. Analysis of driving factors of soil erosion
4.5.1. Using geophysical detectors to analyze the driving factors of soil erosion.
This study selected multiple potential natural and anthropogenic factors influencing soil erosion, including elevation (DEM), slope, normalized difference vegetation index (NDVI), annual precipitation, and land use types. These factors comprehensively represent topography, climate, and land cover, characterized by strong representativeness and reliable data availability. Given the study area’s low population density and limited economic activity, the spatial distribution of GDP lacks representativeness and was therefore excluded from the driver analysis. The factor detector module of the geographic detector was used to evaluate the influence of each factor on the spatial heterogeneity of soil erosion in the Kangding River Basin.
- (1). Factor Detector
The factor detector was employed to explore the driving forces of erosion-related factors in the Kangding River Basin (Table 7). The driving strengths of the factors vary; a smaller q value indicates a weaker influence on the spatial heterogeneity of soil erosion, and vice versa. All five selected factors passed the significance test with p < 0.05. From 2000 to 2010, the order of single-factor q values affecting the soil erosion modulus was: X₂ (slope) > X₅ (vegetation coverage) > X₃ (land use) > X₄ (precipitation) > X₁ (elevation). Among these, X₂ (q = 0.4419) exhibited the strongest driving force, significantly influencing the spatial distribution of soil erosion during this period. In 2015, the ranking changed to: X₃ > X₅ > X₂ > X₄ > X₁, with X₃ (q = 0.4455) becoming the key factor affecting soil erosion. By 2020, the order further shifted to: X₅ > X₃ > X₄ > X₂ > X₁, indicating that X₅ (q = 0.4164) became the dominant driver of soil erosion in the basin during this period.
- (2). Interaction Detector
The results of the geographical detector interaction analysis from 2000 to 2020 are presented in Fig 9. Compared with single factors, the soil erosion variation in the Kangding River Basin exhibits a stronger response to the interaction of any two factors. All interaction detection results fall into the categories of bi-factor enhancement and nonlinear enhancement. In 2000, the interaction between X₃ and other factors was the most prominent, with X₁ ∩ X₃ (q = 0.506) and X₃ ∩ X₅ (q = 0.557) showing the highest explanatory power, indicating that soil erosion that year was significantly driven by the synergistic effects of elevation, land use pattern, and vegetation coverage. In 2005, X₃ ∩ X₅ (q = 0.708) exhibited the strongest interaction effect, followed by X₃ ∩ X₄ (q = 0.606); the combination of land use with vegetation and precipitation became the core drivers of soil erosion. In 2010, X₃ ∩ X₅ (q = 0.680) had the greatest impact on soil erosion, followed by X₁ ∩ X₂ (q = 0.611), indicating significant synergistic effects of vegetation, precipitation, and slope. In 2015, X₃ ∩ X₅ (q = 0.676) and X₃ ∩ X₄ (q = 0.654) remained the leading interactions, with the long-term linkage of land use, vegetation, and precipitation continuously influencing the soil erosion process. By 2020, the q values of various interacting factors tended to balance out, such as X₂ ∩ X₅ (q = 0.501) and X₃ ∩ X₅ (q = 0.560), suggesting enhanced multi-factor synergistic effects among elevation, slope, land use, precipitation, and vegetation coverage. The dominance of single factors weakened, and soil erosion entered a “multi-factor comprehensive driving” stage.
From 2000 to 2020, the driving mechanisms of soil erosion in the Kangding River Basin shifted from being primarily dominated by land use in combination with a single natural factor to a pattern characterized by multi-factor synergy. Soil and water conservation efforts should thus prioritize optimizing land use (e.g., reforestation of farmland, rational utilization of sloped land) and vegetation restoration (enhancing coverage of factor X5), while simultaneously managing the compounded effects of slope and rainfall. Through coordinated multi-factor interventions, the effectiveness of soil erosion control can be significantly improved.
4.5.2. Combining SHAP and machine learning models to analyze the contribution of various factors driving soil erosion.
To compare the simulation performance of different machine learning models for soil erosion intensity, this study employs three modeling methods: CatBoost, XGBoost, and LightGBM. Modeling and validation were conducted for five periods: 2000, 2005, 2010, 2015, and 2020. The best-performing model was selected for SHAP analysis to further investigate the influence mechanisms of various driving factors on soil erosion intensity. Detailed results are presented in Table 8.
The CatBoost model demonstrated the most superior overall performance. Based on the modeling results across five time periods, CatBoost consistently achieved the highest coefficient of determination (R²) in each year, with an average R² of 0.854, significantly outperforming XGBoost (0.794) and LightGBM (0.762). Notably, CatBoost achieved R² values as high as 0.92 and 0.90 in 2010 and 2020, respectively, indicating excellent fitting capability and predictive accuracy. Moreover, the model exhibited relatively small fluctuations in simulation accuracy across the years, reflecting its strong stability and robustness. Therefore, the subsequent analyses of variable importance and spatial interpretation were conducted using the CatBoost model.
To investigate the importance of different driving factors on the model’s predictive outcomes, this study employed the SHAP method based on the CatBoost model to interpret feature importance. Fig 10 illustrates the global SHAP value distributions of the driving factors across the five time periods.
The results indicate that slope had the most significant impact on soil erosion during the first three periods (2000–2010), particularly in 2000, when its SHAP value range was the widest and its contribution far exceeded that of other variables. This finding suggests that during stages dominated by natural driving forces, topography played a decisive role, with steep slopes more prone to concentrated runoff, thereby intensifying topsoil loss. In 2015 and 2020, the importance of Construction land and vegetation coverage increased markedly, becoming the primary driving factors. In 2015, Construction land was dominant, highlighting the profound impact of human disturbance on soil erosion patterns. In 2020, the SHAP value of vegetation coverage surpassed all other variables, becoming the most influential driving factor. As ecological restoration and land management measures have strengthened, the regulatory role of vegetation coverage in erosion processes has gradually increased.
In contrast, precipitation and elevation showed relatively low contributions across all time periods, with consistently small SHAP values and minimal fluctuation. This may be explained by the relatively uniform spatial distribution of precipitation and the limited elevation range in the study area, which reduced their explanatory power compared to other factors.
To reveal the spatial influence characteristics of driving factors on soil erosion, this study utilized SHAP values to generate a spatial heatmap (Fig 11), illustrating the extent to which each factor affects model predictions at the spatial unit level. The x-axis represents the sample indices of spatial locations, while the y-axis corresponds to the five primary driving factors: slope, vegetation coverage, land use, precipitation, and elevation. The color indicates the magnitude and direction of the SHAP value, with red denoting positive contributions (promoting erosion) and blue indicating negative contributions (suppressing erosion). Darker colors signify a stronger influence of the factor on the model output at that grid location. The curve at the top of the figure shows the distribution of predicted values (i.e., the fitted values of the soil erosion modulus), while the lower section displays the corresponding SHAP value matrix, reflecting the strength and direction of each factor’s effect across different spatial locations.
In the SHAP heatmaps across the five time periods, slope and land use were the two factors with the most concentrated contributions to the model output. Slope exhibited high-intensity red areas at multiple spatial locations, indicating a strong positive effect in certain local regions; meanwhile, blue areas also appeared in some locations, suggesting directional variability in its influence on the model results. The SHAP value distribution of land use exhibited clear spatial variation, with alternating red and blue patterns across multiple periods, reflecting strong spatial heterogeneity in its effects.
In most grid cells, vegetation coverage exhibited negative contributions, primarily represented by light to dark blue colors. In certain periods, red responses appeared in specific areas, indicating that its effects were not spatially uniform. The influence of precipitation was more spatially scattered, with some locations showing relatively high SHAP values, generally presenting an intermittent pattern of positive effects.
Elevation showed consistently low SHAP values across the five periods, with most regions displaying no significant contribution and only a few isolated locations showing high values. This indicates that its impact on the model was relatively limited in both magnitude and spatial extent.
The SHAP heatmap revealed spatial differences in sensitivity and effect mechanisms among the driving factors in the Kangding River Basin. Among them, slope and land use exhibited the most prominent spatial contribution patterns, while the inhibitory and promotive effects of vegetation coverage and precipitation varied across regions. These findings highlight the dominant roles of topography and land disturbance in controlling soil erosion and provide a scientific basis for the development of region-specific soil and water conservation strategies.
5. Discussion
To better contextualize the Kangding River Basin within the Upper Yangtze region, its erosion characteristics were compared with two representative basins using unified units (t·hm ⁻ ²·a ⁻ ¹). The upper Minjiang catchment generally showed mild erosion (≈15–16) but with locally intensive zones on steep slopes disturbed by landslides (Bing Guo et al., 2018). The Jialing River Basin, in contrast, was dominated by slight (<5) to light (<25) erosion, exhibiting a steady decline from 1990–2018 while moderate erosion persisted in cultivated mid–lower reaches (Jiang et al., 2025). The Kangding River Basin displayed overall lower absolute values (peak ≈22) but stronger temporal fluctuations—rising sharply in 2000–2005 and then steadily decreasing. Its steep-slope margins resemble the Minjiang Basin, whereas its shift from slope-driven to vegetation-driven control distinguishes it from both basins. These regional contrasts emphasize the unique response of the Kangding River Basin to ecological restoration and land-use policies, providing essential context for the subsequent analysis of driving factors.
5.1. Driving factors of soil erosion changes
5.1.1. Impact of land use change on soil erosion.
Soil erosion is widely recognized as a major environmental threat to terrestrial ecosystems. Identifying its driving factors is therefore essential for effective management. In this study, the geographic detector was applied to evaluate the varying impacts of environmental factors—including elevation, slope, land use, precipitation, and vegetation cover—on soil erosion in the study area. The results indicated that land use change is the primary driving factor of soil erosion in the Kangding River Basin. From 2000 to 2020, the land use types in the basin underwent significant transformations. The proportion of Cropland increased from 2.73% in 2000 to 4.12% in 2020, showing a steady upward trend. These results aligned with patterns showed in the Yellow River Basin, where similar expansion led to increased soil erosion [59]. In the Kangding River Basin, the expansion of Cropland has led to vegetation coverage loss and alterations in soil structure, exacerbating soil erosion and negatively affecting regional ecological balance and environmental sustainability. Similar findings have been reported in the Russian Plain and Ethiopia [60,61]. Meanwhile, Woodland, the dominant land use type in the basin, declined from 60.94% in 2000 to 53.23% in 2020.
According to Wu et al.’s study on vegetation coverage in the western Sichuan Plateau from 1995 to 2015, extensive land degradation occurred in the region due to human disturbances such as deforestation, land reclamation, and tourism development [62]. These activities led to severe soil and water loss on hillside land and a noticeable decline in surface vegetation. Combined with the frequent occurrence of natural disasters in Kangding City in recent years—such as debris flows, landslides, and flash floods—as well as the pressure from tourism development and urban construction [63], the overall trend aligns closely with the findings of this study.
The area of Grassland decreased in 2005 but gradually recovered, reaching 27.07% by 2020. The area of Water body showed a declining trend, decreasing from 1.19% in 2000 to 1.04% in 2020, which poses challenges for ecological balance and water resource management in the basin [64]. Over the past two decades, the area of construction land increased from 2.73% in 2000 to 5.60% in 2020. As a key tourism and cultural center, Kangding City has faced rising land demand driven by urban development and tourism expansion. This has directly contributed to the growth of construction land and highlights the accelerating urbanization process within the Kangding River Basin [65].
5.1.2. Impact of other factors on soil erosion.
Elevation and slope, as fundamental natural elements shaping the topography of the Kangding River Basin, exhibit significant interactions with land use and vegetation coverage. Li et al. applied the geographical detector to analyze the driving factors of soil erosion in the alpine forest-steppe subregion of eastern Qilian Mountains from 2001 to 2020. Their results revealed significant differences in the explanatory power of vegetation cover, elevation, slope, and precipitation, with vegetation cover showing the highest q-value, followed by elevation and slope [66]. Bai’s study on the driving factors of soil erosion in the karst plateau mountainous area of Bijie City from 2000 to 2020 found that slope had the greatest impact, with q-values of 0.283, 0.242, 0.227, 0.159, and 0.336 for the respective years [67], which is consistent with findings in the Kangding River Basin.
Wang evaluated soil erosion in Guiyang City for the years 2008, 2013, and 2018, revealing that the vegetation coverage factor contributed approximately 74% to 92% to the changes in soil erosion, making it the dominant and controlling factor [68]. Similar conclusions were reached by Igwe et al. through studies conducted in Spain, Iran, Algeria, and other region [69].
The precipitation factor is a key driver of runoff changes in river basins, and its spatiotemporal variability directly affects the generation and flow processes as well as the water balance within basins. Meng et al. regarded precipitation as the primary driving force of soil erosion, directly influencing soil particle detachment, soil aggregate breakdown, and the transport of eroded materials [70]. However, in the Kangding River Basin, the contribution of the precipitation factor to runoff and ecological processes was relatively low and did not exhibit significant driving correlations. Its effect was weakened by the basin’s unique underlying surface conditions or other non-precipitation factors. This possibility is also supported by Wang et al., who reported that the contribution of rainfall erosivity factors ranged only from 7.5% to 26.0% [68].
5.1.3. Consistency of factors identified by geographical detector and machine learning model.
To validate the reliability of identifying soil erosion driving factors, this study conducted a comparative analysis between the geographical detector results and the SHAP values derived from the CatBoost model. The single-factor detection results from the geographical detector indicated that slope, land use, and vegetation coverage were the primary driving factors across all periods, exhibiting relatively high q-values. Peak values were observed in 2010 for slope (0.5764), in 2015 for land use (0.4455), and in 2020 for vegetation coverage (0.4164). These peak values highlight the temporal shifts in dominant erosion drivers and reflect the changing balance between natural and anthropogenic influences. The SHAP value analysis from the CatBoost model was highly consistent with these findings.
Between 2000 and 2010, slope remained the dominant factor, with a large dispersion of SHAP values and a significant positive correlation, indicating a sustained positive driving effect of slope on soil erosion. In 2015, land use exhibited the widest SHAP value distribution, where high feature values corresponded to high SHAP values, making it the dominant variable in model output. Such dominance of land use suggests that intensified construction activities and land conversion during this period substantially modified surface conditions, thereby amplifying spatial heterogeneity in soil erosion intensity. In 2020, vegetation coverage had the greatest influence on model predictions, with SHAP values showing a characteristic “low value, high impact” distribution, closely matching the identification results from the geographical detector.
Overall, elevation and precipitation exhibited relatively weak driving effects in both methods, with generally low q-values and SHAP value distributions concentrated near zero. Their limited contributions may be attributed to the relatively uniform spatial distribution of precipitation across the basin and the moderate elevation range, which reduce their explanatory power compared to other factors. The comprehensive comparison demonstrates that the major driving factors identified by the CatBoost model are highly consistent with the results from the geographical detector, providing not only theoretical support and technical assurance for integrated multi-method analyses of soil erosion driving mechanisms but also reinforcing the robustness and credibility of the study’s conclusions.
In summary, slope, land use, and vegetation cover are consistently confirmed as the three dominant driving factors by both methods with different underlying principles, and this cross-validation strengthens the robustness of the conclusion regarding the transition of soil erosion drivers from topography-dominated to human-regulated processes.
5.2. Policies and recommendations
This study analyzed and compared the driving factors of soil erosion in the Kangding River Basin across different periods from 2000 to 2020. The results indicate an overall declining trend in the soil erosion modulus within the basin. However, if the current trends of increasing Cropland area and continuous decrease in Woodland area persist, soil erosion is likely to worsen, leading to higher susceptibility to natural disasters and deterioration of the ecological environment in the basin. Kangding City is one of the cultural and ecological tourism destinations in western China and serves as a core component of the Shangri-La tourism circle spanning Sichuan, Yunnan, and Tibet, as well as an important node in the western Sichuan tourism loop. Tourism development and construction in the study area can damage vegetation coverage and exacerbate soil and water loss. Therefore, enhanced soil monitoring within the basin is imperative, along with the implementation of appropriate conservation and restoration measures. To improve the clarity and applicability of the proposed policies, the recommended measures are briefly summarized into three complementary aspects: planning and regulatory measures, engineering and biotechnological measures, and monitoring and public participation measures.
From the perspective of planning and regulatory measures, the government at all levels in Kangding City should scientifically formulate tourism development plans to avoid excessive development that exceeds the soil’s carrying capacity. It is essential to continue implementing a series of comprehensive ecological environment management policies and measures, such as the Grain-for-Green and Grassland Restoration projects. Currently, Kangding City’s “three key measures” are still being actively promoted (http://www.kangding.gov.cn/, accessed on 2 July 2025). The sequential implementation of these projects has improved vegetation coverage within the basin and effectively curbed soil and water loss as well as landslide hazards, achieving certain positive outcomes. Furthermore, lessons from other Sichuan watersheds suggest the effectiveness of ecological protection red lines in restoring degraded ecosystems. For example, in the Tuojiang and Jialing River basins, the establishment and enforcement of ecological protection red lines, combined with targeted ecological restoration and pollution control measures, have significantly improved water quality, stabilized soil conditions, and enhanced vegetation cover. These cases provide valuable reference points for Kangding River Basin.
In terms of engineering and biotechnological measures, ecological protection red lines should be strictly enforced, and illegal agricultural reclamation and grazing that encroach upon forests and grasslands should be prohibited. Meanwhile, effective soil erosion control should integrate engineering practices such as slope protection and gully treatment with biological approaches, including vegetation restoration, to improve surface stability and soil retention capacity.
Regarding monitoring and public participation measures, a soil erosion monitoring system should be established to timely track erosion dynamics and support adaptive management. In addition, raising environmental awareness among tourists and local residents is crucial for reducing anthropogenic disturbances and promoting eco-tourism that supports coordinated economic and ecological development.
5.3. Limitations and future prospects of the study
Although this study has achieved certain results regarding soil erosion, it has several limitations, primarily due to the relatively small study area. The research focuses on the Kangding River Basin, which, despite exhibiting typical soil erosion characteristics, cannot comprehensively represent the complex conditions of soil erosion influenced by varying geographical environments, climatic conditions, and human activities. Due to the limited scope, the sample size of collected data is constrained. The elevation in the study area generally exceeds 3500 meters, with lower terrain in the eastern part resulting in significant elevation differences. Some locations have slopes reaching up to 40°, leading to poor slope stability. Such topography causes marked spatial heterogeneity in factors like precipitation and runoff within the basin. Consequently, the study does not fully encompass the diverse combinations of topography, vegetation_coverage, and land use types, limiting the generalizability of the findings and posing challenges for applying the conclusions to larger regions.
Secondly, as direct field validation using erosion pins or ^137Cs measurements was not available in this study, the RUSLE-derived soil erosion estimates rely primarily on model calculations and high-resolution remote sensing data. This limitation introduces some uncertainty in the absolute values of soil loss, although the spatial patterns and driving factor analyses remain robust. Such reliance on model-based estimates is a common approach in high-altitude or remote regions where field measurements are difficult to obtain (e.g., mountainous watersheds in western China). Parameter uncertainty arises from measurement and estimation errors in rainfall erosivity (R), soil erodibility (K), slope–length factors (LS), and cover-management (C and P) factors, potentially resulting in overall uncertainties of approximately ±10–20% (refs). Input data precision, including DEM resolution (30 × 30 m) and land-use/vegetation datasets, may limit the detection of microtopographic variations and fine-scale heterogeneity. The RUSLE model also relies on assumptions such as uniform rainfall distribution and simplified soil transport process es, which may not fully capture local erosion dynamics [23,24].
Future research could expand the study area to include multiple ecological regions, ranging from low-elevation hills to high-altitude mountains, enabling comparative analyses of soil erosion characteristics under vertical zonation differences. Incorporating plains with varying wet and dry conditions would facilitate exploration of the effects of precipitation variability on soil erosion. Covering diverse river basins, including the humid and rainy southern basins as well as the arid and drought-prone northern basins, would allow for comprehensive acquisition of rich and diverse datasets. For the complex terrain of the Kangding River Basin, high-resolution remote sensing imagery can be utilized to precisely identify soil erosion conditions across different topographic positions. Coupled with long-term ground monitoring data, this approach would enable analysis of the spatiotemporal evolution of soil erosion under the interactive effects of topography, precipitation, vegetation coverage, and human activities. Moreover, future studies could incorporate field-based validation techniques, such as erosion pins or ^137Cs tracing, to further refine and corroborate RUSLE model estimates. The integration of multi-sensor remote sensing and longer-term monitoring would enhance the reliability of soil erosion assessments under complex terrain and varying climatic conditions. Through multi-regional comparisons, differences in dominant controlling factors of soil erosion under various topographic conditions can be clarified, thereby facilitating the development of more universal soil erosion prediction models that fully consider the critical role of terrain and landform. These improvements will provide a solid foundation for formulating comprehensive and scientifically grounded soil erosion prevention and control strategies tailored to different terrains, effectively addressing the complex and variable challenges of soil erosion in the Kangding River Basin and other similar mountainous watersheds.
6. Conclusion
- Over the past 20 years, the soil erosion modulus in the Kangding River Basin has shown a pattern characterized by an initial increase, followed by a subsequent decrease and eventual stabilization. Between 2000 and 2005, the average soil erosion modulus rose from 16.32 t·hm ⁻ ²·a ⁻ ¹ to a peak of 21.85 t·hm ⁻ ²·a ⁻ ¹, indicating severe soil and water loss during this period. From 2005 onward, the erosion modulus steadily declined, reaching 11.16 t·hm ⁻ ²·a ⁻ ¹ in 2020, the lowest value recorded during the study period. This reflects a significant improvement in soil and water conservation, indicating that the erosion process has become more stable and sustainable.
- The spatial distribution pattern of soil erosion underwent significant changes between 2000 and 2020. Except for a notable increase in erosion area in 2005, the extent of high-intensity zones progressively contracted in other years. Overall, erosion intensity shifted from moderate levels toward light and slight categories, indicating a trend toward more favorable conditions. Nevertheless, areas of slight and light erosion showed considerable fluctuations, with a persistent risk of reverting to moderate or higher levels. Continuous monitoring of these dynamic changes is therefore essential to support soil and water conservation efforts and to prevent localized degradation.
- Temporal analysis from 2000 to 2020 indicates that the dominant controlling factors of soil erosion in the Kangding River Basin have gradually shifted from natural to anthropogenic drivers. Results from the geographical detector reveal that slope was the strongest driving factor in the early period (2000–2010), transitioning to land use dominance by 2015, and finally to vegetation coverage taking the lead in 2020. SHAP analysis further corroborates this trend: slope contributed most to model outputs during the first three periods, subsequently being gradually replaced by land use and vegetation coverage. The high consistency between these two methods suggests that under relatively stable natural conditions, soil erosion has evolved from a “topographic control” type toward an “anthropogenic regulation” type, with ecological restoration and land management measures exerting increasing influence on the erosion process.
References
- 1. Van Oost K, Quine TA, Govers G, De Gryze S, Six J, Harden JW, et al. The impact of agricultural soil erosion on the global carbon cycle. Science. 2007;318(5850):626–9.
- 2. Wei P, Du J, Bahadur A, Zhang H, Wang S, Wu T, et al. Soil erosion and risk assessment on the Qinghai-Tibetan Plateau. Commun Earth Environ. 2025;6(1).
- 3. Borrelli P, Ballabio C, Yang JE, Robinson DA, Panagos P. GloSEM: High-resolution global estimates of present and future soil displacement in croplands by water erosion. Sci Data. 2022;9(1):406. pmid:35831371
- 4. Kong W, Yao Y, Hou L, Wang X, Wei X. Site and landscape position-dependent effects of vegetation removal on soil nitrogen mineralization across five sites on China’s Loess Plateau. CATENA. 2022;215:106336.
- 5.
Mandal D, Roy T. Climate change impact on soil erosion and land degradation. Advances in Global Change Research. Springer Nature Singapore. 2024. 139–61. https://doi.org/10.1007/978-981-99-7935-6_5
- 6. Hussein A. Climate smart agriculture strategies for enhanced agricultural resilience and food security under a changing climate in Ethiopia. Sustainable Environment. 2024;10(1).
- 7. Guo K, Ai G, Ma X, He X. Response of soil erosion to typical topographic factors in the rocky mountain area of North China: Taking Fangshan district as an example. Science of Soil and Water Conservation. 2021;19(4):137–42.
- 8. Bao Pham Q, Ajim Ali S, Parvin F, Van On V, Mohd Sidek L, Đurin B, et al. Multi-spectral remote sensing and GIS-based analysis for decadal land use land cover changes and future prediction using random forest tree and artificial neural network. Advances in Space Research. 2024;74(1):17–47.
- 9. Wen L, Peng Y, Zhou Y, Cai G, Lin Y, Li B. Study on soil erosion and its driving factors from the perspective of landscape in Xiushui watershed, China. Sci Rep. 2023;13(1):8182. pmid:37210377
- 10. Chaudhary S, Wang Y, Dixit AM, Khanal NR, Xu P, Yan K, et al. Eco-environmental risk evaluation for land use planning in areas of potential farmland abandonment in the high mountains of Nepal Himalayas. Sustainability. 2019;11(24):6931.
- 11. Shufang W, Biao Z, Xuejin S. Prediction of soil erosion under different land uses in the typical watershed of the Loess Plateau based on FLUS-CSLE model. Transactions of the Chinese Society of Agricultural Engineering. 2022;38(24):83–92.
- 12. She H, Li P, Wang X, Xiao L, Wang Z, Yang X, et al. Spatial and temporal changes and its driving factors of ecological environment quality in Weihe River basin from 2000 to 2020. Bulletin of Soil and Water Conservation. 2025;45(1):317–26.
- 13. Pan J, Cai F, Yi Z, Zhang W, Yan B, Xue C, et al. Landscape connectivity significantly influences the spatial spillover effects of soil erosion: Based on examples from typical karst watersheds. Ecological Indicators. 2025;173:113373.
- 14. Arunrat N, Sereenonchai S, Kongsurakan P, Hatano R. Soil organic carbon and soil erodibility response to various land-use changes in northern Thailand. CATENA. 2022;219:106595.
- 15. Mokarram M, Zarei AR. Soil erosion prediction using Markov and CA-Markov chains methods and remote sensing drought indicators. Ecological Informatics. 2023;78:102386.
- 16. Seitkazy M, Beisekenov N, Rakhimova M, Tokbergenova A, Zulpykharov K, Kaliyeva D, et al. Soil erosion prediction in western kazakhstan through deep learning with a neural network approach to LS-factor analysis. J Indian Soc Remote Sens. 2024;53(4):1215–26.
- 17. Waseem M, Iqbal F, Humayun M, Umais Latif M, Javed T, Kebede Leta M. Spatial assessment of soil erosion risk using RUSLE Embedded in GIS Environment: A case study of jhelum river watershed. Applied Sciences. 2023;13(6):3775.
- 18. Jiang K, Mo S, Zhang J, Yu K, Li Z. Variation characteristics of soil erosion and their response to landscape patterns in a typical basin in the Upper Yangtze River. Front Plant Sci. 2025;16:1523891. pmid:40182546
- 19. Liu C, Liu J, Zhang L, Shrestha UB, Luo D, Wei Y. Assessing climate and land use change impacts on ecosystem services in the upper Minjiang River basin. Remote Sensing. 2025;17(11):1884.
- 20. Chu L, Sun T, Wang T, Li Z, Cai C. Temporal and spatial heterogeneity of soil erosion and a quantitative analysis of its determinants in the three gorges reservoir area, China. Int J Environ Res Public Health. 2020;17(22):8486. pmid:33207748
- 21. Wang H, Xue H, Yang Y, He W, Liu S, Zhong Y, et al. Multi-Scenario Simulation and Eco-Environmental Effects Analysis of Land Use/Cover Change in China by an Integrated Cellular Automata and Markov Model. Land. 2024;13(4):520.
- 22.
Renard KG. Predicting soil erosion by water: A guide to conservation planning with the Revised Universal Soil Loss equation (RUSLE): Predicting soil erosion by water: A guide to conservation planning with the Revised Universal Soil Loss equation (RUSLE); 1997.
- 23. François M, Gordon CA, Costa de Oliveira U, Rousseau AN, Mariano-Neto E. Assessing the global sensitivity of RUSLE factors: A case study of southern Bahia, Brazil. Soil Systems. 2024;8(4):125.
- 24. Yifan L, Fan Z, Siyi Q, Wenxin L, Desheng C, Shengping W. Seasonal sediment yield estimation on the basis of a revised RUSLE for a typical loess watershed. Science of Soil and Water Conservation. 2025;23(3):73–84.
- 25. Tzioutzios C, Kastridis A. Multi-Criteria Evaluation (MCE) Method for the Management of Woodland Plantations in Floodplain Areas. IJGI. 2020;9(12):725.
- 26. Pangali Sharma TP, Zhang J, Khanal NR, Prodhan FA, Nanzad L, Zhang D, et al. A geomorphic approach for identifying flash flood potential areas in the east rapti river basin of Nepal. IJGI. 2021;10(4):247.
- 27.
Wischmeier WH, Smith DD. Predicting Rainfall Erosion Losses —A Guide To Conservation Planning. USA: United States Dept of Agriculture. 1978.
- 28. Xie Y, Yin S, Liu B, Nearing MA, Zhao Y. Models for estimating daily rainfall erosivity in China. Journal of Hydrology. 2016;535:547–58.
- 29. Chen Y, Xie Y, Duan X, Ding M. Gridded rainfall erosivity (2014–2022) in mainland China using 1 min precipitation data from densely distributed weather stations. Earth Syst Sci Data. 2025;17(3):1265–74.
- 30. Gao G, Liang Y, Liu J, Dunkerley D, Fu B. A modified RUSLE model to simulate soil erosion under different ecological restoration types in the loess hilly area. International Soil and Water Conservation Research. 2024;12(2):258–66.
- 31. Masroor M, Sajjad H, Rehman S, Singh R, Hibjur Rahaman M, Sahana M, et al. Analysing the relationship between drought and soil erosion using vegetation health index and RUSLE models in Godavari middle sub-basin, India. Geoscience Frontiers. 2022;13(2):101312.
- 32. Saha M, Sauda SS, Real HRK, Mahmud M. Estimation of annual rate and spatial distribution of soil erosion in the Jamuna basin using RUSLE model: A geospatial approach. Environmental Challenges. 2022;8:100524.
- 33. Zhang KL, Shu AP, Xu XL, Yang QK, Yu B. Soil erodibility and its estimation for agricultural soils in China. Journal of Arid Environments. 2008;72(6):1002–11.
- 34. Labrière N, Locatelli B, Laumonier Y, Freycon V, Bernoux M. Soil erosion in the humid tropics: A systematic quantitative review. Agriculture, Ecosystems & Environment. 2015;203:127–39.
- 35. Molla T, Sisheber B. Estimating soil erosion risk and evaluating erosion control measures for soil conservation planning at Koga watershed in the highlands of Ethiopia. Solid Earth. 2017;8(1):13–25.
- 36. Ganasri BP, Ramesh H. Assessment of soil erosion by RUSLE model using remote sensing and GIS - A case study of Nethravathi Basin. Geoscience Frontiers. 2016;7(6):953–61.
- 37. Chongfa C, Shuwen D, Zhihua S, Li H, Guangyuan Z. Study of applying USLE and geographical information system IDRISI to predict soil erosion in small watershed. Journal of Soil and Water Conservation. 2000;14(2):19–24.
- 38. Li H, Guan Q, Sun Y, Wang Q, Liang L, Ma Y, et al. Spatiotemporal analysis of the quantitative attribution of soil water erosion in the upper reaches of the Yellow River Basin based on the RUSLE-TLSD model. CATENA. 2022;212:106081.
- 39. Alitane A, Essahlaoui A, El Hafyani M, El Hmaidi A, El Ouali A, Kassou A, et al. Water erosion monitoring and prediction in response to the effects of climate change using RUSLE and SWAT equations: Case of R’Dom Watershed in Morocco. Land. 2022;11(1):93.
- 40. Guo X, Gui X, Xiong H, Hu X, Li Y, Cui H, et al. Critical role of climate factors for groundwater potential mapping in arid regions: Insights from random forest, XGBoost, and LightGBM algorithms. Journal of Hydrology. 2023;621:129599.
- 41.
Kangding Municipal People’s Government. Bulletin on the Main Data of the Third National Land Survey of Kangding City Kangding, Sichuan, China2022 http://www.kangding.gov.cn/kdsrmzf/c104217/202209/1f52f0ed659042679f9dd5ed840fce62.shtml
- 42. Wang S, Xu X, Huang L. Spatial and temporal variability of soil erosion in northeast China from 2000 to 2020. Remote Sensing. 2023;15(1).
- 43. Dou X, Ma X, Zhao C, Li J, Yan Y, Zhu J. Risk assessment of soil erosion in Central Asia under global warming. CATENA. 2022;212:106056.
- 44. Fang H, Guo M. Aspect-induced differences in soil erosion intensity in a gullied hilly region on the Chinese Loess Plateau. Environ Earth Sci. 2015;74(7):5677–85.
- 45. Qiao X, Li Z, Lin J, Wang H, Zheng S, Yang S. Assessing current and future soil erosion under changing land use based on InVEST and FLUS models in the Yihe River Basin, North China. International Soil and Water Conservation Research. 2024;12(2):298–312.
- 46. Stefanidis S, Alexandridis V, Ghosal K. Assessment of Water-Induced Soil Erosion as a Threat to Natura 2000 Protected Areas in Crete Island, Greece. Sustainability. 2022;14(5):2738.
- 47. Tian P, Tian X, Geng R, Zhao G, Yang L, Mu X, et al. Response of soil erosion to vegetation restoration and terracing on the Loess Plateau. CATENA. 2023;227:107103.
- 48. Borrelli P, Panagos P, Alewell C, Ballabio C, de Oliveira Fagundes H, Haregeweyn N, et al. Policy implications of multiple concurrent soil erosion processes in European farmland. Nat Sustain. 2022;6(1):103–12.
- 49. Jinfeng W. Geoprobes: Principles and Prospects. Acta Geographica Sinica. 2017;72(1):116–34.
- 50. Gao Y, Wang L, Zhong G, Wang Y, Yang J. Potential of remote sensing images for soil moisture retrieving using ensemble learning methods in vegetation-covered area. IEEE J Sel Top Appl Earth Observations Remote Sensing. 2023;16:8149–65.
- 51.
Lundberg SM, Lee SI. Proceedings of the 31st International Conference on Neural Information Processing Systems, Long Beach, California, USA, 2017. 4768–77.
- 52. Wang H, Liang Q, Hancock JT, Khoshgoftaar TM. Feature selection strategies: A comparative analysis of SHAP-value and importance-based methods. J Big Data. 2024;11(1).
- 53.
Marcilio WE, Eler DM. From explanations to feature selection: Assessing SHAP values as feature selection mechanism. 2020 33rd SIBGRAPI Conference on Graphics, Patterns and Images (SIBGRAPI), 2020. 340–7. https://doi.org/10.1109/sibgrapi51738.2020.00053
- 54. Pawley SM, Atkinson L, Utting DJ, Hartman GMD, Atkinson N. Evaluating spatially enabled machine learning approaches to depth to bedrock mapping, Alberta, Canada. PLoS One. 2024;19(3):e0296881. pmid:38536867
- 55. Regmi NR, Webb NDS, Walter JI, Heo J, Hayman NW. Mapping landforms of a hilly landscape using machine learning and high-resolution LiDAR topographic data. Applied Computing and Geosciences. 2024;24:100203.
- 56. Pan S, Zheng Z, Guo Z, Luo H. An optimized XGBoost method for predicting reservoir porosity using petrophysical logs. Journal of Petroleum Science and Engineering. 2022;208:109520.
- 57. Li Z. Extracting spatial effects from machine learning model using local interpretation method: An example of SHAP and XGBoost. Computers, Environment and Urban Systems. 2022;96:101845.
- 58. Ma M, Zhao G, He B, Li Q, Dong H, Wang S, et al. XGBoost-based method for flash flood risk assessment. Journal of Hydrology. 2021;598:126382.
- 59.
Hao C. Spatial and temporal changes and driving factors of soil erosion in watersheds before and after returning farmland to forest on the Loess Plateau. Xianyang: North West Agriculture and Forestry University. 2019.
- 60. Golosov V, Yermolaev OP, Rysin II, Litvin L, Kiryukhina ZP, Safina GR. Influence of climate and land use changes on recent trend of soil erosion within the Russian Plain. European geosciences union general assembly. 2016;18:6569.
- 61. Adimassu Z, Tamene L, Degefie DT. The influence of grazing and cultivation on runoff, soil erosion, and soil nutrient export in the central highlands of Ethiopia. Ecol Process. 2020;9(1).
- 62. Liangxu W, Han W, Huaiyong S, Sitong Z. Spatiotemporal pattern of vegetation and its response to climate change in the western Sichuan Plateau. Research of Soil and Water Conservation. 2021;28(1):171–8.
- 63. Feng W, Fan Y, Zhongrong J, e W, Guan W. Susceptibility assessment of debris flow based on watershed units in Kangding City, Sichuan Province. The Chinese Journal of Geological Hazard and Control. 2023;34(3):145–56.
- 64. Gao J, Jiang Y, Wang H, Zuo L. Identification of Dominant Factors Affecting Soil Erosion and Water Yield within Ecological Red Line Areas. Remote Sensing. 2020;12(3):399.
- 65. Zhang J, Jin L, Pan X, Wang Y. Pro-Environmental Behavior of Tourists in Ecotourism Scenic Spots: The Promoting Role of Tourist Experience Quality in Place Attachment. Sustainability. 2024;16(20):8984.
- 66. Li M, Liu X, Ding J, Zhang W, Wang R, Luo X, et al. Analysis of soil erosion dynamics and its driving factors in the Qilian Mountains of Qingdong. Soil Research. 2024;62(6).
- 67. Bai Y, Zhang Y, Zhang S, Wu J, Zhao X, Zhao F. Characteristics and driving factors of spatiotemporal changes in soil erosion in the karst plateau mountainous region over 20 years. PLoS One. 2024;19(11):e0314266. pmid:39570933
- 68. Zhijie W, Shujun L, Yuan S. Spatial-Temporal Characteristies of Soil Erosion and Erosion Hotspots in Guiyang City of Karst Plateau Mountainous Region in 2008-2018. Journal of Soil and Water Conservation. 2020;34(5):94–110.
- 69. Igwe PU, Ezeukwu JC, Edoka NE, Ejie OC, Ifi GI. A review of vegetation cover as a natural factor to soil erosion. International Journal of Rural Development, Environment and Health Research. 2017;1(4):21–8.
- 70. Meng X, Zhu Y, Yin M, Liu D. The impact of land use and rainfall patterns on the soil loss of the hillslope. Sci Rep. 2021;11(1):16341. pmid:34381135